Skip to content

Latest commit

 

History

History
327 lines (253 loc) · 12.6 KB

CHANGELOG.rst

File metadata and controls

327 lines (253 loc) · 12.6 KB
Raw

Changelog

0.12

0.12.2

  • improved the downsampling method of output variables of the Network.run method - now, an average over all simulated samples is calculated for the downsampling
  • updated documentation and use examples to account for the recent changes to the Network.add_diffeq_node and Network.run methods
  • updated pytests to account for new downsampling method

0.12.1

  • added a cutoff keyword argument to the Network.run function that allows to define the number of initial steps that should be disregarded in the simulation results
  • updated PyPI config for uploading new versions

0.12.0

  • added option to define spiking networks which contain the spike reset mechanism in their governing differential equations
  • changed the name of the keyword argument spike_def of the method Network.add_diffeq_node to reset_var
  • added a keyword argument reset to the Network.add_diffeq_node that controls whether the spike reset mechanism of RectiPy should be used for a given node or whether it has an intrinsic spike reset mechanism
  • added a QIF neuron template that uses an intrinsic spike reset mechanism
  • updated the pytest library to account for the changes
  • dropped official support for python 3.6 and added support for python 3.10

0.11

0.11.1

  • updated readthedocs configuration file
  • added new node classes MemoryNode and MemoryNet that allow to implement delay coupling in networks
  • renamed the FunctionNode class into InstantNode to better set it apart from other node types that have some type of intrinsic memory
  • improved method for detaching tensors via Network.detach() - provides more control over which tensors to detach now.
  • debugged FeedbackNetwork.compile() method to work with multiple calls of the method
  • introduced edge masking as a new feature. Can be used by passing a mask array to the keyword argument mask of Network.add_edge

0.11.0

  • added a class FeedbackNetwork that allows to add feedback connections between nodes in a network. Intended use: First, build the feedforward network structure that maps from input to output. Then add feedback edges, marking them via the keyword argument feedback=True of the FeedbackNetwork.add_edge method.
  • both feedforward and feedback edges can be trained
  • Re-worked the backpropagation through time and recursive least squares training methods to work with feedback weights
  • updated the readthedocs documentation
  • added new pytests for the different optimization methods: Ridge regression, backpropagation, and recursive least squares
  • added a class MultiSpikeNode that allows to define multiple spiking variables within a single differential equation-based node.

0.10

0.10.3

  • updated documentation of user interfaces
  • updated use examples for user interfaces
  • removed a minor bug in nodes.py that led to issues with retrieving variable indices from single-neuron nodes
  • removed a bug in nodes.py where, in some cases, the state vector y was not moved to the correct device after state resets

0.10.2

  • debugged the Network.reset() method. Properly resets the network to the passed state dictionary now, if provided.
  • improved the RateNet.reset() method. The method can now handle both numpy arrays as well as tensors as input.

0.10.1

  • added a new method Network.clear() that removes all nodes and edges from a rectipy.Network instance
  • added new keyword arguments to Network.detach() that allow to customize the behavior of the method with respect to which tensors are going to be detached and whether they should require a gradient after detachment or not
  • added a new method Network.set_var() that allows to change the values of variables in the network instance
  • added a new method Node.set_param() that allows to change the values of parameters of the node
  • improved documentation of the methods of rectipy.Network

0.10.0

  • Changed user interface: rectipy.Network is now initialized as an empty graph. Nodes and edges can be added to it afternwards.
  • Added new modules: rectipy.nodes and rectipy.edges provide different classes for nodes and edges that can be added to rectipy.Network instances
  • Altered training functionalities: rectipy.Network now features two methods for parameter optimization: Network.fit_ridge for Ridge regression-based training of a set of readout weights, and Network.fit_bptt for gradient descent based on backpropagation through time.
  • Implemented truncated backpropagation through time
  • Implemented gradient surrogates for spiking neural networks
  • New functionalities of the rectipy.Observer module: All recorded outputs can now be returned either as a list of torch.Tensor objects, as numpy arrays, or as pandas.DataFrame objects.
  • Added convenience functions on rectipy.Network for adding and getting nodes, edges and network variables.
  • Improved integration of rectipy.Network with pytorch parameter optimization methods by adding high-level methods for (i) detaching all state-variables from the current graph for gradient computation, and (ii) resetting the state of the entire network.
  • Updated unit tests and documentation to work with the above described changes
  • Updated PyRates interface to work with recent changes to the pyrates.CircuitTemplate.add_edges_from_matrix method

0.9

0.9.3

  • added a new IK neuron template with biexponential synaptic dynamics
  • improved layout of the readthedocs documentation website
  • changed the readout function to use the SGDClassifier instead of Ridge from sklearn

0.9.2

  • cleaned some code after model deployment changes made in 0.9.1
  • debugged the from_template initialization methods

0.9.1

  • minor bug fix of faulty normalization of input weights in utility.input_connections
  • resolved issues with the model deployment on a certain device. Instead of providing the device ("cpu" or "cuda") to the Network.compile method, it is to be provided during initialization now
  • debugged network initialization method Network.from_template

0.9.0

  • debugged global recovery variable definition of izhikevich model template
  • debugged simulation test
  • added a new rectipy.Network initialization method: Network.from_template that allows to initialize Network instances from pyrates.CircuitTemplate instances. This way, the user has full control over the construction of the network template.

0.8

0.8.0

  • added a use example for rectipy-torch integration
  • added a function for matrix normalization to utility
  • added the izhikevich neuron model as a template
  • added an izhikevich neuron with global recovery variable as a template

0.7

0.7.0

  • added visualization method rectipy.observer.Observer.matshow that allows to create 2D color-coded plots of multi-dimensional RNN state variables
  • simplified alteration of default parameter values during network initialization
  • added use example for training and testing via the Network.train and Network.test methods
  • added a global coupling constant k to the qif model template
  • improved docstrings

0.6

0.6.0

  • added use example for the LIF neuron model
  • new variable views available on the rectipy.Network and rectipy.rnn_layer.RNNLayer classes
  • Network.__getitem__() and RNNLayer.__getitem__() allow to directly access parameters and variables of the RNNLayer instance
  • integrated the new variable views into the documentation and testing suite
  • simplified code for model definitions based on the new variable views

0.5

0.5.2

  • added use example for the QIF neuron models
  • added use example for the leaky-integrator rate neuron model
  • added use example gallery skeleton
  • added use example for network initialization
  • added use example for numerical simulations
  • added use example for the observer
  • removed bug from SRNNLayer that caused model initialization to fail when no dtype for variales was provided
  • removed bug from the sigmoid operator that is part of the leaky_integrator.yaml model definition file
  • added .gitignore file
  • added model template for LIF neurons
  • improved docstrings of the Network class

0.5.1

  • added documentation source files for a readthedocs documentation website
  • added yaml configuration and config files for readthedocs installation
  • added a first use example
  • added installation instructions
  • added the changelog to the readthedocs website sources
  • added a full API section
  • renamed the tests module to rectipy_tests to avoid confusion with the PyRates.tests module

0.5.0

  • reduced overhead of InputLayer and OutputLayer by making them return instances of torch.nn.Linear or rectipy.input_layer.LinearStatic upon initialization
  • reduced overhead of Network.compile by directly accessing the torch.Module instances to create the torch.Sequential
  • improved test library with more extensive testing of RNNLayer and Network functionalities

0.4

0.4.1

  • added new pytests that test the functionalities of the RNNLayer.record and RNNLayer.reset methods
  • added new pytests that test the initialization functions of Network
  • improved integration of PyRates into RectiPy, by making sure that all PyRates caches are cleared, even if building the network functions fails due to erroneous user inputs

0.4.0

  • removed all in-place operations for non-spiking networks
  • changed pyrates interface such that vector-field updates are not performed in-place anymore
  • only in-place operation left: Spike resetting
  • added methods Network.forward and Network.parameters that allow the class Network to be embedded in larger network structures.
  • added method RNNLayer.reset as a method that can be used to reset the state vector of the RNN
  • added new tests for the rnn layer
  • debugged detach method in rnn layer
  • debugged issues with in-place operations and autograd
  • added a new example for parameter fitting within the RNN layer

0.3

0.3.1

  • improved documentation
  • added pytests for the initialization functions of the rnn layer
  • debugged index-finding functions for trainable parameters in the rnn layer
  • improved integration of pyrates functions into rnn layer

0.3.0

  • added utility function readout that allows to train a readout classifier on collected network states and targets
  • added new gradient descent optimizer options
  • added possibility of making an optimizer step only every x training steps (gradients will accumulate over these steps)

0.2

0.2.0

  • renamed the model template package to avoid interference with the pyrates-intrinsic model template package
  • added a utility function for the generation of input weight matrices
  • added a utility function for winner-takes-all score calculation
  • added getitem methods to the Network (integer-based indexing, returns layers) and Observer (string-based indexing, returns recordings) classes
  • added the possibility to the Network.train method to train in epochs
  • made the device argument of Network.compile optional
  • ensured that the activation functions of the OutputLayer are always applied to the first dimension of the outputs

0.1

0.1.5

  • ensured that state variable indices in RNN layer use correct data type (torch.int64)

0.1.4

  • added pytests for the output layer
  • added checks on the correctness of the input arguments for the output layer
  • added keyword arguments to the OutputLayer.__init__() that are passed on to torch.nn.Linear if trainable=True

0.1.3

  • added pytests for the input layer
  • added a CircleCI config
  • added automated execution of all tests via CircleCI upon pushing to github
  • added pytest to the requirements

0.1.2

  • added docstrings to the Network class for all non-private methods
  • added docstrings to the Obsever class for all non-private methods
  • made Network.compile a public method and reduced the number of automatized calls to it by Network (Network.train, Network.test and Network.run only call Network.compile themselves if it hasn’t been done before)
  • added a public property Network.model that provides read access to the pytorch model of the network

0.1.1

  • added automated pypi releases
  • added github workflow for pypi releases
  • updated readme

0.1.0

  • code structure:
    • network class as main user interface
    • input, output, and rnn layers as network components
    • observer as class for results storage
  • model templates package for yaml definition files
  • installation instructions